This document will be used to keep track of changes made between release versions. I'll do my best to note any breaking changes!

• sinhrks

• None

• Add a new datasets module behind a datasets feature flag.
• Add new classification scores: precision, recall, and f1.
• Add a new Transformer::fit function to allow prefitting of a Transformer before use.

• None

• LinRegressor now uses solve instead of inverse for improved accuracy and stability.

• None

• Adding a new confusion_matrix module.

• None

• Updated rulinalg dependency to 0.3.7.

• scholtzan

• None

• None

• Regularization constant for GMM is now only added to diagonal.

• Added some better Result handling to GMM.

This version includes no changes but is a bump due to a crates bug.

See the notes for 0.5.0 below.

This is another fairly large release. Thank you to everyone who contributed!

• NivenT
• theotherphil
• andrewcsmith

• The SupModel and UnSupModel traits now return Results for the train and predict functions.
• Updated to rulinalg v0.3 (see rulinalg changelog for details).

• Adding RMSProp gradient descent algorithm. #121
• Adding cross validation. #125
• Adding a new Shuffler transformer. #135

• None

• Adding benchmarks
• Initiate GMM with sample covariance of data (instead of identity matrix).

• None

• Adding new Transformer trait for data preprocessing.
• Adding a MinMax transformer.
• Adding a Standardizer transformer.

• None

• tafia who is responsible for all changes in this release.

• None

• None

• Made neural nets more efficient by reducing clones and some restructuring.
• Removing unneeded copying in favour of slicing for performance.
• Using iter_rows in favour of manually row iterating by chunks.

• None

• None

• Fixed a significant bug in the K-Means algorithm. Centroids were not updating correctly during M-step.

• None

• Added experimental implementation of DBSCAN clustering.

• Added new example for K-Means clustering in repo.

This is the biggest release so far. Primarily because the linalg module has been pulled out into its own crate: rulinalg.

In addition to this there have been a number of improvements to the linalg and learning moduled in this release.

• The linalg module pulled out and replaced by reexports of rulinalg. All structs are now imported at the linalg level, i.e. linalg::matrix::Matrix -> linalg::Matrix.
• Decomposition methods now return Result instead of panicking on fail.
• K-Means now has a trait for Initializer - which allows generic initialization algorithms.

• New error handling in both the linalg (now rulinalg) and learning modules.
• Bug fixed in eigendecomposition: it can now be used!
• K-means can now take a generic initialization algorithm.

• Optimization and code cleanup in the decomposition methods.
• Some optimization in the K-Means model.

• ic (Added examples to repo!)

• Parameter methods now return Option<&Type> instead of &Option<Type>.

• MatrixSlice and MatrixSliceMut now have IntoIterator methods.

• Adding examples to the repository.

• DarkDrek (Who is responsible for almost all changes in this release. Thank you!)

• Matrix: mean and variance methods now take Axes enum instead of usize flag for dimension.

• Assignment operators (+=, -=, etc.) now implemented for Vector.

• Some optimizations to variance computation for Matrix.
• Some code cleanup - thanks to clippy.

• None

• New helper methods to access GMM distribution parameters.
• New GMM constructor to choose different prior mixture weights.

• Fixed a bug where GMM covariances were incorrectly computed when using diagonal constraint.

• rrichardson

• All fields on GradDesc and StochasticGD are now private.
• Matrix slices now have the same lifetime as their target data.

• Adding new slice utility methods : from_raw_parts for MatrixSlices and as_slice methods for Matrix.
• Adding framework for regularization. Implementing regularization for nnets.
• Adding early stopping to gradient descent algorithms.
• Adding AdaGrad gradient descent algorithm.
• Implementing Into and From for Matrix, Vector, and MatrixSlices.

• Bug fixing naive bayes : no longer attempts to update empty class.
• Removing unneeded trait bounds on Matrix/Vector implementations.

• The new constructors for Matrix and Vector now take an Into<Vec> generic type. May break some type inference.

• Added row iterators for each matrix struct.
• Implemented OpAssign overloading for Matrix and MatrixSliceMut.

• Moved unit tests into respective modules.
• Modified slice iterators to make the offset usage safe(er).
• Removed some compiler warnings from the tests.

• None

• Matrix and Vector now implement PartialEq.

• Fixed a bug where eigendecomposition for 2x2 matrices was incorrect.

• None

• None

• Fixing a bug with matrix slice multiplication.
• Removing unneeded NumCast import.

• None

• Adding Naive Bayes classifiers.
• Adding a prelude for common imports.
• Adding MatrixSlice and MatrixSliceMut for efficient matrix views.

• Using matrixmultiply to get huge performance gains! Thanks bluss.
• Code refactor to split up the matrix module.

• vishalsodani (fixing some typos)
• danlrobertson (added the KMeansClassifierBuilder)

• None

• KMeansClassifier now has a builder!

• We're now using travis for CI.
• Deriving Debug, Clone, Copy for Gaussian and Exponential distributions.

• mut_data method now returns a mutable slice &mut [T] instead of a Vec<T>.

• More vectorization and optimization of linear algebra.

• Copy and Clone now implemented where applicable.
• Added test coverage.

• zackmdavis (contributed all features for this version, thank you!)

• None

• Can now debug print matrices and vectors.
• Can now pretty print matrices to given precision.

• Fixed the dependency versions used in Cargo.toml.
• Updated the library documentation with complete list of ML tools.

• None

• Addition of Gaussian Mixture Models.
• Allow basic arithmetic to combine kernels.

• Added some missing documentation.
• Some code formatting.
• Minor improvements thanks to clippy.

• Neural network instantiation new method now requires a training algorithm to be specified.

• Adding more kernels (for full list see API documentation).
• Generalized Linear Model.
• Updated model structures to allow more freedom in training algorithms.

• Some more documentation.
• Some minor code formatting.

• None

• Add Support Vector Machines.

• Minor code cleanup.
• Some micro optimization.

• None

• Added the stats module behind the optional feature flag stats.
• stats currently includes support for the Exponential and Gaussian distributions.

• Some rustfmt code cleanup.

• Removed the new constructor for the LinRegressor. This has been replaced by the default function from the Default trait.

• Added a select method for cloning a block from a matrix.
• Implemented QR decomposition, and eigenvalue decomposition.
• Implemented eigendecomp (though only works definitely for real-symmetric matrices).

• Optimizations to matrix multiplication